Figure 1.
Data acquisition and handling oversight.
Data was collected from four different sources (see Methods). The first panel shows how many pMHCs were derived from each data set and their respective MHC restrictions and immunogenicity status. Data from all sets was combined, the number of non-redundant 9mers with respect to the host in which the data was obtained is shown in the second panel.
Figure 2.
T-cell preferences for different amino acids in HLA class I presented peptides.
The fraction of an amino acid in immunogenic (left bar, filled) and non-immunogenic (right bar, unfilled) peptides presented on HLA class I molecules is shown. Significantly different distributions are indicated with a star (Permutation test, see Methods: p<0.05; False discovery rate (FDR) for multiple testing determined as in [46]: q<0.05). The background frequency for each amino acid in the protein sequences that were a source of the immunogenic or non-immunogenic peptides is shown by a grey line.
Table 1.
Amino acid characteristics of immunogenic peptides presented on HLA class I molecules.
Table 2.
Position dependent differences between immunogenic and non-immunogenic peptides.
Figure 3.
Cross-validation of the immunogenicity model.
Two-thirds of the data were used for making the immunogenicity model (see methods) and one-third for cross-validation. The average ROC (thick grey line) of 25 of such cross-validations (thin lines) are plotted. The average AUC was 0.65.
Figure 4.
Predicting Dengue-derived CTL epitopes with the immunogenicity model.
Immunogenicity scores were determined for non-redundant epitopes (n = 22) and non-epitopes (n = 110) identified in mice by Weiskopf et al. [55] (A), and for non-redundant epitopes (n = 42) and non-epitopes (n = 477) identified by Weiskopf et al. in humans [56] (B). Average and variation of the average are shown as thick lines with error bars, individual scores are shown as dots. Both in mice and in men, the epitopes had a significantly higher immunogenicity score than the non-epitopes (Murine data (A): p<0.01 (Wilcoxon rank-sum test); AUC = 0.69. Human data (B): p = 0.014 (Wilcoxon rank-sum test); AUC = 0.61).
Figure 5.
Viral pMHCs are better recognized by T-cells.
For common HLA molecules (13 HLA-A and 15 HLA-B), viral and human ligands were predicted using MHC binding and precursor protein processing predictors (as in [40]). The fraction of viral pMHCs (y-axis) and human pMHCs (x-axis) with a positive score in our immunogenicity model is shown. The diagonal denotes the line y = x, HLA molecules with a larger fraction of positively scoring viral pMHCs fall above this line, which was the case for 27 of the 28 HLA molecules (sign-test: p<0.001). Three HLA molecules where the difference between the (predicted) viral and human ligands was largest (B*0702, A*3001, B*4501), and one HLA molecule where the difference between viral and human ligands was smallest (A*2301), and HLA-A*0201 are indicated in the figure.